drivers/hv/channel_mgmt.c

   1 /*
   2  * Copyright (c) 2009, Microsoft Corporation.
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms and conditions of the GNU General Public License,
   6  * version 2, as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  11  * more details.
  12  *
  13  * You should have received a copy of the GNU General Public License along with
  14  * this program; if not, write to the Free Software Foundation, Inc., 59 Temple
  15  * Place - Suite 330, Boston, MA 02111-1307 USA.
  16  *
  17  * Authors:
  18  *   Haiyang Zhang <haiyangz@microsoft.com>
  19  *   Hank Janssen  <hjanssen@microsoft.com>
  20  */
  21 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  22
  23 #include <linux/kernel.h>
  24 #include <linux/interrupt.h>
  25 #include <linux/sched.h>
  26 #include <linux/wait.h>
  27 #include <linux/mm.h>
  28 #include <linux/slab.h>
  29 #include <linux/list.h>
  30 #include <linux/module.h>
  31 #include <linux/completion.h>
  32 #include <linux/delay.h>
  33 #include <linux/hyperv.h>
  34 #include <asm/mshyperv.h>
  35
  36 #include "hyperv_vmbus.h"
  37
  38 static void init_vp_index(struct vmbus_channel *channel, u16 dev_type);
  39
  40 static const struct vmbus_device vmbus_devs[] = {
  41         /* IDE */
  42         { .dev_type = HV_IDE,
  43           HV_IDE_GUID,
  44           .perf_device = true,
  45         },
  46
  47         /* SCSI */
  48         { .dev_type = HV_SCSI,
  49           HV_SCSI_GUID,
  50           .perf_device = true,
  51         },
  52
  53         /* Fibre Channel */
  54         { .dev_type = HV_FC,
  55           HV_SYNTHFC_GUID,
  56           .perf_device = true,
  57         },
  58
  59         /* Synthetic NIC */
  60         { .dev_type = HV_NIC,
  61           HV_NIC_GUID,
  62           .perf_device = true,
  63         },
  64
  65         /* Network Direct */
  66         { .dev_type = HV_ND,
  67           HV_ND_GUID,
  68           .perf_device = true,
  69         },
  70
  71         /* PCIE */
  72         { .dev_type = HV_PCIE,
  73           HV_PCIE_GUID,
  74           .perf_device = true,
  75         },
  76
  77         /* Synthetic Frame Buffer */
  78         { .dev_type = HV_FB,
  79           HV_SYNTHVID_GUID,
  80           .perf_device = false,
  81         },
  82
  83         /* Synthetic Keyboard */
  84         { .dev_type = HV_KBD,
  85           HV_KBD_GUID,
  86           .perf_device = false,
  87         },
  88
  89         /* Synthetic MOUSE */
  90         { .dev_type = HV_MOUSE,
  91           HV_MOUSE_GUID,
  92           .perf_device = false,
  93         },
  94
  95         /* KVP */
  96         { .dev_type = HV_KVP,
  97           HV_KVP_GUID,
  98           .perf_device = false,
  99         },
 100
 101         /* Time Synch */
 102         { .dev_type = HV_TS,
 103           HV_TS_GUID,
 104           .perf_device = false,
 105         },
 106
 107         /* Heartbeat */
 108         { .dev_type = HV_HB,
 109           HV_HEART_BEAT_GUID,
 110           .perf_device = false,
 111         },
 112
 113         /* Shutdown */
 114         { .dev_type = HV_SHUTDOWN,
 115           HV_SHUTDOWN_GUID,
 116           .perf_device = false,
 117         },
 118
 119         /* File copy */
 120         { .dev_type = HV_FCOPY,
 121           HV_FCOPY_GUID,
 122           .perf_device = false,
 123         },
 124
 125         /* Backup */
 126         { .dev_type = HV_BACKUP,
 127           HV_VSS_GUID,
 128           .perf_device = false,
 129         },
 130
 131         /* Dynamic Memory */
 132         { .dev_type = HV_DM,
 133           HV_DM_GUID,
 134           .perf_device = false,
 135         },
 136
 137         /* Unknown GUID */
 138         { .dev_type = HV_UNKNOWN,
 139           .perf_device = false,
 140         },
 141 };
 142
 143 static const struct {
 144         uuid_le guid;
 145 } vmbus_unsupported_devs[] = {
 146         { HV_AVMA1_GUID },
 147         { HV_AVMA2_GUID },
 148         { HV_RDV_GUID   },
 149 };
 150
 151 /*
 152  * The rescinded channel may be blocked waiting for a response from the host;
 153  * take care of that.
 154  */
 155 static void vmbus_rescind_cleanup(struct vmbus_channel *channel)
 156 {
 157         struct vmbus_channel_msginfo *msginfo;
 158         unsigned long flags;
 159
 160
 161         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 162
 163         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
 164                                 msglistentry) {
 165
 166                 if (msginfo->waiting_channel == channel) {
 167                         complete(&msginfo->waitevent);
 168                         break;
 169                 }
 170         }
 171         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
 172 }
 173
 174 static bool is_unsupported_vmbus_devs(const uuid_le *guid)
 175 {
 176         int i;
 177
 178         for (i = 0; i < ARRAY_SIZE(vmbus_unsupported_devs); i++)
 179                 if (!uuid_le_cmp(*guid, vmbus_unsupported_devs[i].guid))
 180                         return true;
 181         return false;
 182 }
 183
 184 static u16 hv_get_dev_type(const struct vmbus_channel *channel)
 185 {
 186         const uuid_le *guid = &channel->offermsg.offer.if_type;
 187         u16 i;
 188
 189         if (is_hvsock_channel(channel) || is_unsupported_vmbus_devs(guid))
 190                 return HV_UNKNOWN;
 191
 192         for (i = HV_IDE; i < HV_UNKNOWN; i++) {
 193                 if (!uuid_le_cmp(*guid, vmbus_devs[i].guid))
 194                         return i;
 195         }
 196         pr_info("Unknown GUID: %pUl\n", guid);
 197         return i;
 198 }
 199
 200 /**
 201  * vmbus_prep_negotiate_resp() - Create default response for Hyper-V Negotiate message
 202  * @icmsghdrp: Pointer to msg header structure
 203  * @icmsg_negotiate: Pointer to negotiate message structure
 204  * @buf: Raw buffer channel data
 205  *
 206  * @icmsghdrp is of type &struct icmsg_hdr.
 207  * Set up and fill in default negotiate response message.
 208  *
 209  * The fw_version and fw_vercnt specifies the framework version that
 210  * we can support.
 211  *
 212  * The srv_version and srv_vercnt specifies the service
 213  * versions we can support.
 214  *
 215  * Versions are given in decreasing order.
 216  *
 217  * nego_fw_version and nego_srv_version store the selected protocol versions.
 218  *
 219  * Mainly used by Hyper-V drivers.
 220  */
 221 bool vmbus_prep_negotiate_resp(struct icmsg_hdr *icmsghdrp,
 222                                 u8 *buf, const int *fw_version, int fw_vercnt,
 223                                 const int *srv_version, int srv_vercnt,
 224                                 int *nego_fw_version, int *nego_srv_version)
 225 {
 226         int icframe_major, icframe_minor;
 227         int icmsg_major, icmsg_minor;
 228         int fw_major, fw_minor;
 229         int srv_major, srv_minor;
 230         int i, j;
 231         bool found_match = false;
 232         struct icmsg_negotiate *negop;
 233
 234         icmsghdrp->icmsgsize = 0x10;
 235         negop = (struct icmsg_negotiate *)&buf[
 236                 sizeof(struct vmbuspipe_hdr) +
 237                 sizeof(struct icmsg_hdr)];
 238
 239         icframe_major = negop->icframe_vercnt;
 240         icframe_minor = 0;
 241
 242         icmsg_major = negop->icmsg_vercnt;
 243         icmsg_minor = 0;
 244
 245         /*
 246          * Select the framework version number we will
 247          * support.
 248          */
 249
 250         for (i = 0; i < fw_vercnt; i++) {
 251                 fw_major = (fw_version[i] >> 16);
 252                 fw_minor = (fw_version[i] & 0xFFFF);
 253
 254                 for (j = 0; j < negop->icframe_vercnt; j++) {
 255                         if ((negop->icversion_data[j].major == fw_major) &&
 256                             (negop->icversion_data[j].minor == fw_minor)) {
 257                                 icframe_major = negop->icversion_data[j].major;
 258                                 icframe_minor = negop->icversion_data[j].minor;
 259                                 found_match = true;
 260                                 break;
 261                         }
 262                 }
 263
 264                 if (found_match)
 265                         break;
 266         }
 267
 268         if (!found_match)
 269                 goto fw_error;
 270
 271         found_match = false;
 272
 273         for (i = 0; i < srv_vercnt; i++) {
 274                 srv_major = (srv_version[i] >> 16);
 275                 srv_minor = (srv_version[i] & 0xFFFF);
 276
 277                 for (j = negop->icframe_vercnt;
 278                         (j < negop->icframe_vercnt + negop->icmsg_vercnt);
 279                         j++) {
 280
 281                         if ((negop->icversion_data[j].major == srv_major) &&
 282                                 (negop->icversion_data[j].minor == srv_minor)) {
 283
 284                                 icmsg_major = negop->icversion_data[j].major;
 285                                 icmsg_minor = negop->icversion_data[j].minor;
 286                                 found_match = true;
 287                                 break;
 288                         }
 289                 }
 290
 291                 if (found_match)
 292                         break;
 293         }
 294
 295         /*
 296          * Respond with the framework and service
 297          * version numbers we can support.
 298          */
 299
 300 fw_error:
 301         if (!found_match) {
 302                 negop->icframe_vercnt = 0;
 303                 negop->icmsg_vercnt = 0;
 304         } else {
 305                 negop->icframe_vercnt = 1;
 306                 negop->icmsg_vercnt = 1;
 307         }
 308
 309         if (nego_fw_version)
 310                 *nego_fw_version = (icframe_major << 16) | icframe_minor;
 311
 312         if (nego_srv_version)
 313                 *nego_srv_version = (icmsg_major << 16) | icmsg_minor;
 314
 315         negop->icversion_data[0].major = icframe_major;
 316         negop->icversion_data[0].minor = icframe_minor;
 317         negop->icversion_data[1].major = icmsg_major;
 318         negop->icversion_data[1].minor = icmsg_minor;
 319         return found_match;
 320 }
 321
 322 EXPORT_SYMBOL_GPL(vmbus_prep_negotiate_resp);
 323
 324 /*
 325  * alloc_channel - Allocate and initialize a vmbus channel object
 326  */
 327 static struct vmbus_channel *alloc_channel(void)
 328 {
 329         struct vmbus_channel *channel;
 330
 331         channel = kzalloc(sizeof(*channel), GFP_ATOMIC);
 332         if (!channel)
 333                 return NULL;
 334
 335         spin_lock_init(&channel->inbound_lock);
 336         spin_lock_init(&channel->lock);
 337
 338         INIT_LIST_HEAD(&channel->sc_list);
 339         INIT_LIST_HEAD(&channel->percpu_list);
 340
 341         tasklet_init(&channel->callback_event,
 342                      vmbus_on_event, (unsigned long)channel);
 343
 344         return channel;
 345 }
 346
 347 /*
 348  * free_channel - Release the resources used by the vmbus channel object
 349  */
 350 static void free_channel(struct vmbus_channel *channel)
 351 {
 352         tasklet_kill(&channel->callback_event);
 353         kfree(channel);
 354 }
 355
 356 static void percpu_channel_enq(void *arg)
 357 {
 358         struct vmbus_channel *channel = arg;
 359         struct hv_per_cpu_context *hv_cpu
 360                 = this_cpu_ptr(hv_context.cpu_context);
 361
 362         list_add_tail(&channel->percpu_list, &hv_cpu->chan_list);
 363 }
 364
 365 static void percpu_channel_deq(void *arg)
 366 {
 367         struct vmbus_channel *channel = arg;
 368
 369         list_del(&channel->percpu_list);
 370 }
 371
 372
 373 static void vmbus_release_relid(u32 relid)
 374 {
 375         struct vmbus_channel_relid_released msg;
 376
 377         memset(&msg, 0, sizeof(struct vmbus_channel_relid_released));
 378         msg.child_relid = relid;
 379         msg.header.msgtype = CHANNELMSG_RELID_RELEASED;
 380         vmbus_post_msg(&msg, sizeof(struct vmbus_channel_relid_released),
 381                        true);
 382 }
 383
 384 void hv_event_tasklet_disable(struct vmbus_channel *channel)
 385 {
 386         tasklet_disable(&channel->callback_event);
 387 }
 388
 389 void hv_event_tasklet_enable(struct vmbus_channel *channel)
 390 {
 391         tasklet_enable(&channel->callback_event);
 392
 393         /* In case there is any pending event */
 394         tasklet_schedule(&channel->callback_event);
 395 }
 396
 397 void hv_process_channel_removal(struct vmbus_channel *channel, u32 relid)
 398 {
 399         unsigned long flags;
 400         struct vmbus_channel *primary_channel;
 401
 402         BUG_ON(!channel->rescind);
 403         BUG_ON(!mutex_is_locked(&vmbus_connection.channel_mutex));
 404
 405         hv_event_tasklet_disable(channel);
 406         if (channel->target_cpu != get_cpu()) {
 407                 put_cpu();
 408                 smp_call_function_single(channel->target_cpu,
 409                                          percpu_channel_deq, channel, true);
 410         } else {
 411                 percpu_channel_deq(channel);
 412                 put_cpu();
 413         }
 414         hv_event_tasklet_enable(channel);
 415
 416         if (channel->primary_channel == NULL) {
 417                 list_del(&channel->listentry);
 418
 419                 primary_channel = channel;
 420         } else {
 421                 primary_channel = channel->primary_channel;
 422                 spin_lock_irqsave(&primary_channel->lock, flags);
 423                 list_del(&channel->sc_list);
 424                 primary_channel->num_sc--;
 425                 spin_unlock_irqrestore(&primary_channel->lock, flags);
 426         }
 427
 428         /*
 429          * We need to free the bit for init_vp_index() to work in the case
 430          * of sub-channel, when we reload drivers like hv_netvsc.
 431          */
 432         if (channel->affinity_policy == HV_LOCALIZED)
 433                 cpumask_clear_cpu(channel->target_cpu,
 434                                   &primary_channel->alloced_cpus_in_node);
 435
 436         vmbus_release_relid(relid);
 437
 438         free_channel(channel);
 439 }
 440
 441 void vmbus_free_channels(void)
 442 {
 443         struct vmbus_channel *channel, *tmp;
 444
 445         mutex_lock(&vmbus_connection.channel_mutex);
 446         list_for_each_entry_safe(channel, tmp, &vmbus_connection.chn_list,
 447                 listentry) {
 448                 /* hv_process_channel_removal() needs this */
 449                 channel->rescind = true;
 450
 451                 vmbus_device_unregister(channel->device_obj);
 452         }
 453         mutex_unlock(&vmbus_connection.channel_mutex);
 454 }
 455
 456 /*
 457  * vmbus_process_offer - Process the offer by creating a channel/device
 458  * associated with this offer
 459  */
 460 static void vmbus_process_offer(struct vmbus_channel *newchannel)
 461 {
 462         struct vmbus_channel *channel;
 463         bool fnew = true;
 464         unsigned long flags;
 465         u16 dev_type;
 466         int ret;
 467
 468         /* Make sure this is a new offer */
 469         mutex_lock(&vmbus_connection.channel_mutex);
 470
 471         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
 472                 if (!uuid_le_cmp(channel->offermsg.offer.if_type,
 473                         newchannel->offermsg.offer.if_type) &&
 474                         !uuid_le_cmp(channel->offermsg.offer.if_instance,
 475                                 newchannel->offermsg.offer.if_instance)) {
 476                         fnew = false;
 477                         break;
 478                 }
 479         }
 480
 481         if (fnew)
 482                 list_add_tail(&newchannel->listentry,
 483                               &vmbus_connection.chn_list);
 484
 485         mutex_unlock(&vmbus_connection.channel_mutex);
 486
 487         if (!fnew) {
 488                 /*
 489                  * Check to see if this is a sub-channel.
 490                  */
 491                 if (newchannel->offermsg.offer.sub_channel_index != 0) {
 492                         /*
 493                          * Process the sub-channel.
 494                          */
 495                         newchannel->primary_channel = channel;
 496                         spin_lock_irqsave(&channel->lock, flags);
 497                         list_add_tail(&newchannel->sc_list, &channel->sc_list);
 498                         channel->num_sc++;
 499                         spin_unlock_irqrestore(&channel->lock, flags);
 500                 } else
 501                         goto err_free_chan;
 502         }
 503
 504         dev_type = hv_get_dev_type(newchannel);
 505
 506         init_vp_index(newchannel, dev_type);
 507
 508         hv_event_tasklet_disable(newchannel);
 509         if (newchannel->target_cpu != get_cpu()) {
 510                 put_cpu();
 511                 smp_call_function_single(newchannel->target_cpu,
 512                                          percpu_channel_enq,
 513                                          newchannel, true);
 514         } else {
 515                 percpu_channel_enq(newchannel);
 516                 put_cpu();
 517         }
 518         hv_event_tasklet_enable(newchannel);
 519
 520         /*
 521          * This state is used to indicate a successful open
 522          * so that when we do close the channel normally, we
 523          * can cleanup properly
 524          */
 525         newchannel->state = CHANNEL_OPEN_STATE;
 526
 527         if (!fnew) {
 528                 if (channel->sc_creation_callback != NULL)
 529                         channel->sc_creation_callback(newchannel);
 530                 return;
 531         }
 532
 533         /*
 534          * Start the process of binding this offer to the driver
 535          * We need to set the DeviceObject field before calling
 536          * vmbus_child_dev_add()
 537          */
 538         newchannel->device_obj = vmbus_device_create(
 539                 &newchannel->offermsg.offer.if_type,
 540                 &newchannel->offermsg.offer.if_instance,
 541                 newchannel);
 542         if (!newchannel->device_obj)
 543                 goto err_deq_chan;
 544
 545         newchannel->device_obj->device_id = dev_type;
 546         /*
 547          * Add the new device to the bus. This will kick off device-driver
 548          * binding which eventually invokes the device driver's AddDevice()
 549          * method.
 550          */
 551         mutex_lock(&vmbus_connection.channel_mutex);
 552         ret = vmbus_device_register(newchannel->device_obj);
 553         mutex_unlock(&vmbus_connection.channel_mutex);
 554
 555         if (ret != 0) {
 556                 pr_err("unable to add child device object (relid %d)\n",
 557                         newchannel->offermsg.child_relid);
 558                 kfree(newchannel->device_obj);
 559                 goto err_deq_chan;
 560         }
 561         return;
 562
 563 err_deq_chan:
 564         mutex_lock(&vmbus_connection.channel_mutex);
 565         list_del(&newchannel->listentry);
 566         mutex_unlock(&vmbus_connection.channel_mutex);
 567
 568         hv_event_tasklet_disable(newchannel);
 569         if (newchannel->target_cpu != get_cpu()) {
 570                 put_cpu();
 571                 smp_call_function_single(newchannel->target_cpu,
 572                                          percpu_channel_deq, newchannel, true);
 573         } else {
 574                 percpu_channel_deq(newchannel);
 575                 put_cpu();
 576         }
 577         hv_event_tasklet_enable(newchannel);
 578
 579         vmbus_release_relid(newchannel->offermsg.child_relid);
 580
 581 err_free_chan:
 582         free_channel(newchannel);
 583 }
 584
 585 /*
 586  * We use this state to statically distribute the channel interrupt load.
 587  */
 588 static int next_numa_node_id;
 589
 590 /*
 591  * Starting with Win8, we can statically distribute the incoming
 592  * channel interrupt load by binding a channel to VCPU.
 593  * We do this in a hierarchical fashion:
 594  * First distribute the primary channels across available NUMA nodes
 595  * and then distribute the subchannels amongst the CPUs in the NUMA
 596  * node assigned to the primary channel.
 597  *
 598  * For pre-win8 hosts or non-performance critical channels we assign the
 599  * first CPU in the first NUMA node.
 600  */
 601 static void init_vp_index(struct vmbus_channel *channel, u16 dev_type)
 602 {
 603         u32 cur_cpu;
 604         bool perf_chn = vmbus_devs[dev_type].perf_device;
 605         struct vmbus_channel *primary = channel->primary_channel;
 606         int next_node;
 607         struct cpumask available_mask;
 608         struct cpumask *alloced_mask;
 609
 610         if ((vmbus_proto_version == VERSION_WS2008) ||
 611             (vmbus_proto_version == VERSION_WIN7) || (!perf_chn)) {
 612                 /*
 613                  * Prior to win8, all channel interrupts are
 614                  * delivered on cpu 0.
 615                  * Also if the channel is not a performance critical
 616                  * channel, bind it to cpu 0.
 617                  */
 618                 channel->numa_node = 0;
 619                 channel->target_cpu = 0;
 620                 channel->target_vp = hv_context.vp_index[0];
 621                 return;
 622         }
 623
 624         /*
 625          * Based on the channel affinity policy, we will assign the NUMA
 626          * nodes.
 627          */
 628
 629         if ((channel->affinity_policy == HV_BALANCED) || (!primary)) {
 630                 while (true) {
 631                         next_node = next_numa_node_id++;
 632                         if (next_node == nr_node_ids) {
 633                                 next_node = next_numa_node_id = 0;
 634                                 continue;
 635                         }
 636                         if (cpumask_empty(cpumask_of_node(next_node)))
 637                                 continue;
 638                         break;
 639                 }
 640                 channel->numa_node = next_node;
 641                 primary = channel;
 642         }
 643         alloced_mask = &hv_context.hv_numa_map[primary->numa_node];
 644
 645         if (cpumask_weight(alloced_mask) ==
 646             cpumask_weight(cpumask_of_node(primary->numa_node))) {
 647                 /*
 648                  * We have cycled through all the CPUs in the node;
 649                  * reset the alloced map.
 650                  */
 651                 cpumask_clear(alloced_mask);
 652         }
 653
 654         cpumask_xor(&available_mask, alloced_mask,
 655                     cpumask_of_node(primary->numa_node));
 656
 657         cur_cpu = -1;
 658
 659         if (primary->affinity_policy == HV_LOCALIZED) {
 660                 /*
 661                  * Normally Hyper-V host doesn't create more subchannels
 662                  * than there are VCPUs on the node but it is possible when not
 663                  * all present VCPUs on the node are initialized by guest.
 664                  * Clear the alloced_cpus_in_node to start over.
 665                  */
 666                 if (cpumask_equal(&primary->alloced_cpus_in_node,
 667                                   cpumask_of_node(primary->numa_node)))
 668                         cpumask_clear(&primary->alloced_cpus_in_node);
 669         }
 670
 671         while (true) {
 672                 cur_cpu = cpumask_next(cur_cpu, &available_mask);
 673                 if (cur_cpu >= nr_cpu_ids) {
 674                         cur_cpu = -1;
 675                         cpumask_copy(&available_mask,
 676                                      cpumask_of_node(primary->numa_node));
 677                         continue;
 678                 }
 679
 680                 if (primary->affinity_policy == HV_LOCALIZED) {
 681                         /*
 682                          * NOTE: in the case of sub-channel, we clear the
 683                          * sub-channel related bit(s) in
 684                          * primary->alloced_cpus_in_node in
 685                          * hv_process_channel_removal(), so when we
 686                          * reload drivers like hv_netvsc in SMP guest, here
 687                          * we're able to re-allocate
 688                          * bit from primary->alloced_cpus_in_node.
 689                          */
 690                         if (!cpumask_test_cpu(cur_cpu,
 691                                               &primary->alloced_cpus_in_node)) {
 692                                 cpumask_set_cpu(cur_cpu,
 693                                                 &primary->alloced_cpus_in_node);
 694                                 cpumask_set_cpu(cur_cpu, alloced_mask);
 695                                 break;
 696                         }
 697                 } else {
 698                         cpumask_set_cpu(cur_cpu, alloced_mask);
 699                         break;
 700                 }
 701         }
 702
 703         channel->target_cpu = cur_cpu;
 704         channel->target_vp = hv_context.vp_index[cur_cpu];
 705 }
 706
 707 static void vmbus_wait_for_unload(void)
 708 {
 709         int cpu;
 710         void *page_addr;
 711         struct hv_message *msg;
 712         struct vmbus_channel_message_header *hdr;
 713         u32 message_type;
 714
 715         /*
 716          * CHANNELMSG_UNLOAD_RESPONSE is always delivered to the CPU which was
 717          * used for initial contact or to CPU0 depending on host version. When
 718          * we're crashing on a different CPU let's hope that IRQ handler on
 719          * the cpu which receives CHANNELMSG_UNLOAD_RESPONSE is still
 720          * functional and vmbus_unload_response() will complete
 721          * vmbus_connection.unload_event. If not, the last thing we can do is
 722          * read message pages for all CPUs directly.
 723          */
 724         while (1) {
 725                 if (completion_done(&vmbus_connection.unload_event))
 726                         break;
 727
 728                 for_each_online_cpu(cpu) {
 729                         struct hv_per_cpu_context *hv_cpu
 730                                 = per_cpu_ptr(hv_context.cpu_context, cpu);
 731
 732                         page_addr = hv_cpu->synic_message_page;
 733                         msg = (struct hv_message *)page_addr
 734                                 + VMBUS_MESSAGE_SINT;
 735
 736                         message_type = READ_ONCE(msg->header.message_type);
 737                         if (message_type == HVMSG_NONE)
 738                                 continue;
 739
 740                         hdr = (struct vmbus_channel_message_header *)
 741                                 msg->u.payload;
 742
 743                         if (hdr->msgtype == CHANNELMSG_UNLOAD_RESPONSE)
 744                                 complete(&vmbus_connection.unload_event);
 745
 746                         vmbus_signal_eom(msg, message_type);
 747                 }
 748
 749                 mdelay(10);
 750         }
 751
 752         /*
 753          * We're crashing and already got the UNLOAD_RESPONSE, cleanup all
 754          * maybe-pending messages on all CPUs to be able to receive new
 755          * messages after we reconnect.
 756          */
 757         for_each_online_cpu(cpu) {
 758                 struct hv_per_cpu_context *hv_cpu
 759                         = per_cpu_ptr(hv_context.cpu_context, cpu);
 760
 761                 page_addr = hv_cpu->synic_message_page;
 762                 msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
 763                 msg->header.message_type = HVMSG_NONE;
 764         }
 765 }
 766
 767 /*
 768  * vmbus_unload_response - Handler for the unload response.
 769  */
 770 static void vmbus_unload_response(struct vmbus_channel_message_header *hdr)
 771 {
 772         /*
 773          * This is a global event; just wakeup the waiting thread.
 774          * Once we successfully unload, we can cleanup the monitor state.
 775          */
 776         complete(&vmbus_connection.unload_event);
 777 }
 778
 779 void vmbus_initiate_unload(bool crash)
 780 {
 781         struct vmbus_channel_message_header hdr;
 782
 783         /* Pre-Win2012R2 hosts don't support reconnect */
 784         if (vmbus_proto_version < VERSION_WIN8_1)
 785                 return;
 786
 787         init_completion(&vmbus_connection.unload_event);
 788         memset(&hdr, 0, sizeof(struct vmbus_channel_message_header));
 789         hdr.msgtype = CHANNELMSG_UNLOAD;
 790         vmbus_post_msg(&hdr, sizeof(struct vmbus_channel_message_header),
 791                        !crash);
 792
 793         /*
 794          * vmbus_initiate_unload() is also called on crash and the crash can be
 795          * happening in an interrupt context, where scheduling is impossible.
 796          */
 797         if (!crash)
 798                 wait_for_completion(&vmbus_connection.unload_event);
 799         else
 800                 vmbus_wait_for_unload();
 801 }
 802
 803 /*
 804  * vmbus_onoffer - Handler for channel offers from vmbus in parent partition.
 805  *
 806  */
 807 static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
 808 {
 809         struct vmbus_channel_offer_channel *offer;
 810         struct vmbus_channel *newchannel;
 811
 812         offer = (struct vmbus_channel_offer_channel *)hdr;
 813
 814         /* Allocate the channel object and save this offer. */
 815         newchannel = alloc_channel();
 816         if (!newchannel) {
 817                 pr_err("Unable to allocate channel object\n");
 818                 return;
 819         }
 820
 821         /*
 822          * Setup state for signalling the host.
 823          */
 824         newchannel->sig_event = (struct hv_input_signal_event *)
 825                                 (ALIGN((unsigned long)
 826                                 &newchannel->sig_buf,
 827                                 HV_HYPERCALL_PARAM_ALIGN));
 828
 829         newchannel->sig_event->connectionid.asu32 = 0;
 830         newchannel->sig_event->connectionid.u.id = VMBUS_EVENT_CONNECTION_ID;
 831         newchannel->sig_event->flag_number = 0;
 832         newchannel->sig_event->rsvdz = 0;
 833
 834         if (vmbus_proto_version != VERSION_WS2008) {
 835                 newchannel->is_dedicated_interrupt =
 836                                 (offer->is_dedicated_interrupt != 0);
 837                 newchannel->sig_event->connectionid.u.id =
 838                                 offer->connection_id;
 839         }
 840
 841         memcpy(&newchannel->offermsg, offer,
 842                sizeof(struct vmbus_channel_offer_channel));
 843         newchannel->monitor_grp = (u8)offer->monitorid / 32;
 844         newchannel->monitor_bit = (u8)offer->monitorid % 32;
 845
 846         vmbus_process_offer(newchannel);
 847 }
 848
 849 /*
 850  * vmbus_onoffer_rescind - Rescind offer handler.
 851  *
 852  * We queue a work item to process this offer synchronously
 853  */
 854 static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
 855 {
 856         struct vmbus_channel_rescind_offer *rescind;
 857         struct vmbus_channel *channel;
 858         unsigned long flags;
 859         struct device *dev;
 860
 861         rescind = (struct vmbus_channel_rescind_offer *)hdr;
 862
 863         mutex_lock(&vmbus_connection.channel_mutex);
 864         channel = relid2channel(rescind->child_relid);
 865
 866         if (channel == NULL) {
 867                 /*
 868                  * This is very impossible, because in
 869                  * vmbus_process_offer(), we have already invoked
 870                  * vmbus_release_relid() on error.
 871                  */
 872                 goto out;
 873         }
 874
 875         spin_lock_irqsave(&channel->lock, flags);
 876         channel->rescind = true;
 877         spin_unlock_irqrestore(&channel->lock, flags);
 878
 879         vmbus_rescind_cleanup(channel);
 880
 881         if (channel->device_obj) {
 882                 if (channel->chn_rescind_callback) {
 883                         channel->chn_rescind_callback(channel);
 884                         goto out;
 885                 }
 886                 /*
 887                  * We will have to unregister this device from the
 888                  * driver core.
 889                  */
 890                 dev = get_device(&channel->device_obj->device);
 891                 if (dev) {
 892                         vmbus_device_unregister(channel->device_obj);
 893                         put_device(dev);
 894                 }
 895         } else {
 896                 hv_process_channel_removal(channel,
 897                         channel->offermsg.child_relid);
 898         }
 899
 900 out:
 901         mutex_unlock(&vmbus_connection.channel_mutex);
 902 }
 903
 904 void vmbus_hvsock_device_unregister(struct vmbus_channel *channel)
 905 {
 906         mutex_lock(&vmbus_connection.channel_mutex);
 907
 908         BUG_ON(!is_hvsock_channel(channel));
 909
 910         channel->rescind = true;
 911         vmbus_device_unregister(channel->device_obj);
 912
 913         mutex_unlock(&vmbus_connection.channel_mutex);
 914 }
 915 EXPORT_SYMBOL_GPL(vmbus_hvsock_device_unregister);
 916
 917
 918 /*
 919  * vmbus_onoffers_delivered -
 920  * This is invoked when all offers have been delivered.
 921  *
 922  * Nothing to do here.
 923  */
 924 static void vmbus_onoffers_delivered(
 925                         struct vmbus_channel_message_header *hdr)
 926 {
 927 }
 928
 929 /*
 930  * vmbus_onopen_result - Open result handler.
 931  *
 932  * This is invoked when we received a response to our channel open request.
 933  * Find the matching request, copy the response and signal the requesting
 934  * thread.
 935  */
 936 static void vmbus_onopen_result(struct vmbus_channel_message_header *hdr)
 937 {
 938         struct vmbus_channel_open_result *result;
 939         struct vmbus_channel_msginfo *msginfo;
 940         struct vmbus_channel_message_header *requestheader;
 941         struct vmbus_channel_open_channel *openmsg;
 942         unsigned long flags;
 943
 944         result = (struct vmbus_channel_open_result *)hdr;
 945
 946         /*
 947          * Find the open msg, copy the result and signal/unblock the wait event
 948          */
 949         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 950
 951         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
 952                                 msglistentry) {
 953                 requestheader =
 954                         (struct vmbus_channel_message_header *)msginfo->msg;
 955
 956                 if (requestheader->msgtype == CHANNELMSG_OPENCHANNEL) {
 957                         openmsg =
 958                         (struct vmbus_channel_open_channel *)msginfo->msg;
 959                         if (openmsg->child_relid == result->child_relid &&
 960                             openmsg->openid == result->openid) {
 961                                 memcpy(&msginfo->response.open_result,
 962                                        result,
 963                                        sizeof(
 964                                         struct vmbus_channel_open_result));
 965                                 complete(&msginfo->waitevent);
 966                                 break;
 967                         }
 968                 }
 969         }
 970         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
 971 }
 972
 973 /*
 974  * vmbus_ongpadl_created - GPADL created handler.
 975  *
 976  * This is invoked when we received a response to our gpadl create request.
 977  * Find the matching request, copy the response and signal the requesting
 978  * thread.
 979  */
 980 static void vmbus_ongpadl_created(struct vmbus_channel_message_header *hdr)
 981 {
 982         struct vmbus_channel_gpadl_created *gpadlcreated;
 983         struct vmbus_channel_msginfo *msginfo;
 984         struct vmbus_channel_message_header *requestheader;
 985         struct vmbus_channel_gpadl_header *gpadlheader;
 986         unsigned long flags;
 987
 988         gpadlcreated = (struct vmbus_channel_gpadl_created *)hdr;
 989
 990         /*
 991          * Find the establish msg, copy the result and signal/unblock the wait
 992          * event
 993          */
 994         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 995
 996         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
 997                                 msglistentry) {
 998                 requestheader =
 999                         (struct vmbus_channel_message_header *)msginfo->msg;
1000
1001                 if (requestheader->msgtype == CHANNELMSG_GPADL_HEADER) {
1002                         gpadlheader =
1003                         (struct vmbus_channel_gpadl_header *)requestheader;
1004
1005                         if ((gpadlcreated->child_relid ==
1006                              gpadlheader->child_relid) &&
1007                             (gpadlcreated->gpadl == gpadlheader->gpadl)) {
1008                                 memcpy(&msginfo->response.gpadl_created,
1009                                        gpadlcreated,
1010                                        sizeof(
1011                                         struct vmbus_channel_gpadl_created));
1012                                 complete(&msginfo->waitevent);
1013                                 break;
1014                         }
1015                 }
1016         }
1017         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1018 }
1019
1020 /*
1021  * vmbus_ongpadl_torndown - GPADL torndown handler.
1022  *
1023  * This is invoked when we received a response to our gpadl teardown request.
1024  * Find the matching request, copy the response and signal the requesting
1025  * thread.
1026  */
1027 static void vmbus_ongpadl_torndown(
1028                         struct vmbus_channel_message_header *hdr)
1029 {
1030         struct vmbus_channel_gpadl_torndown *gpadl_torndown;
1031         struct vmbus_channel_msginfo *msginfo;
1032         struct vmbus_channel_message_header *requestheader;
1033         struct vmbus_channel_gpadl_teardown *gpadl_teardown;
1034         unsigned long flags;
1035
1036         gpadl_torndown = (struct vmbus_channel_gpadl_torndown *)hdr;
1037
1038         /*
1039          * Find the open msg, copy the result and signal/unblock the wait event
1040          */
1041         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1042
1043         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1044                                 msglistentry) {
1045                 requestheader =
1046                         (struct vmbus_channel_message_header *)msginfo->msg;
1047
1048                 if (requestheader->msgtype == CHANNELMSG_GPADL_TEARDOWN) {
1049                         gpadl_teardown =
1050                         (struct vmbus_channel_gpadl_teardown *)requestheader;
1051
1052                         if (gpadl_torndown->gpadl == gpadl_teardown->gpadl) {
1053                                 memcpy(&msginfo->response.gpadl_torndown,
1054                                        gpadl_torndown,
1055                                        sizeof(
1056                                         struct vmbus_channel_gpadl_torndown));
1057                                 complete(&msginfo->waitevent);
1058                                 break;
1059                         }
1060                 }
1061         }
1062         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1063 }
1064
1065 /*
1066  * vmbus_onversion_response - Version response handler
1067  *
1068  * This is invoked when we received a response to our initiate contact request.
1069  * Find the matching request, copy the response and signal the requesting
1070  * thread.
1071  */
1072 static void vmbus_onversion_response(
1073                 struct vmbus_channel_message_header *hdr)
1074 {
1075         struct vmbus_channel_msginfo *msginfo;
1076         struct vmbus_channel_message_header *requestheader;
1077         struct vmbus_channel_version_response *version_response;
1078         unsigned long flags;
1079
1080         version_response = (struct vmbus_channel_version_response *)hdr;
1081         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1082
1083         list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1084                                 msglistentry) {
1085                 requestheader =
1086                         (struct vmbus_channel_message_header *)msginfo->msg;
1087
1088                 if (requestheader->msgtype ==
1089                     CHANNELMSG_INITIATE_CONTACT) {
1090                         memcpy(&msginfo->response.version_response,
1091                               version_response,
1092                               sizeof(struct vmbus_channel_version_response));
1093                         complete(&msginfo->waitevent);
1094                 }
1095         }
1096         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1097 }
1098
1099 /* Channel message dispatch table */
1100 struct vmbus_channel_message_table_entry
1101         channel_message_table[CHANNELMSG_COUNT] = {
1102         {CHANNELMSG_INVALID,                    0, NULL},
1103         {CHANNELMSG_OFFERCHANNEL,               0, vmbus_onoffer},
1104         {CHANNELMSG_RESCIND_CHANNELOFFER,       0, vmbus_onoffer_rescind},
1105         {CHANNELMSG_REQUESTOFFERS,              0, NULL},
1106         {CHANNELMSG_ALLOFFERS_DELIVERED,        1, vmbus_onoffers_delivered},
1107         {CHANNELMSG_OPENCHANNEL,                0, NULL},
1108         {CHANNELMSG_OPENCHANNEL_RESULT,         1, vmbus_onopen_result},
1109         {CHANNELMSG_CLOSECHANNEL,               0, NULL},
1110         {CHANNELMSG_GPADL_HEADER,               0, NULL},
1111         {CHANNELMSG_GPADL_BODY,                 0, NULL},
1112         {CHANNELMSG_GPADL_CREATED,              1, vmbus_ongpadl_created},
1113         {CHANNELMSG_GPADL_TEARDOWN,             0, NULL},
1114         {CHANNELMSG_GPADL_TORNDOWN,             1, vmbus_ongpadl_torndown},
1115         {CHANNELMSG_RELID_RELEASED,             0, NULL},
1116         {CHANNELMSG_INITIATE_CONTACT,           0, NULL},
1117         {CHANNELMSG_VERSION_RESPONSE,           1, vmbus_onversion_response},
1118         {CHANNELMSG_UNLOAD,                     0, NULL},
1119         {CHANNELMSG_UNLOAD_RESPONSE,            1, vmbus_unload_response},
1120         {CHANNELMSG_18,                         0, NULL},
1121         {CHANNELMSG_19,                         0, NULL},
1122         {CHANNELMSG_20,                         0, NULL},
1123         {CHANNELMSG_TL_CONNECT_REQUEST,         0, NULL},
1124 };
1125
1126 /*
1127  * vmbus_onmessage - Handler for channel protocol messages.
1128  *
1129  * This is invoked in the vmbus worker thread context.
1130  */
1131 void vmbus_onmessage(void *context)
1132 {
1133         struct hv_message *msg = context;
1134         struct vmbus_channel_message_header *hdr;
1135         int size;
1136
1137         hdr = (struct vmbus_channel_message_header *)msg->u.payload;
1138         size = msg->header.payload_size;
1139
1140         if (hdr->msgtype >= CHANNELMSG_COUNT) {
1141                 pr_err("Received invalid channel message type %d size %d\n",
1142                            hdr->msgtype, size);
1143                 print_hex_dump_bytes("", DUMP_PREFIX_NONE,
1144                                      (unsigned char *)msg->u.payload, size);
1145                 return;
1146         }
1147
1148         if (channel_message_table[hdr->msgtype].message_handler)
1149                 channel_message_table[hdr->msgtype].message_handler(hdr);
1150         else
1151                 pr_err("Unhandled channel message type %d\n", hdr->msgtype);
1152 }
1153
1154 /*
1155  * vmbus_request_offers - Send a request to get all our pending offers.
1156  */
1157 int vmbus_request_offers(void)
1158 {
1159         struct vmbus_channel_message_header *msg;
1160         struct vmbus_channel_msginfo *msginfo;
1161         int ret;
1162
1163         msginfo = kmalloc(sizeof(*msginfo) +
1164                           sizeof(struct vmbus_channel_message_header),
1165                           GFP_KERNEL);
1166         if (!msginfo)
1167                 return -ENOMEM;
1168
1169         msg = (struct vmbus_channel_message_header *)msginfo->msg;
1170
1171         msg->msgtype = CHANNELMSG_REQUESTOFFERS;
1172
1173
1174         ret = vmbus_post_msg(msg, sizeof(struct vmbus_channel_message_header),
1175                              true);
1176         if (ret != 0) {
1177                 pr_err("Unable to request offers - %d\n", ret);
1178
1179                 goto cleanup;
1180         }
1181
1182 cleanup:
1183         kfree(msginfo);
1184
1185         return ret;
1186 }
1187
1188 /*
1189  * Retrieve the (sub) channel on which to send an outgoing request.
1190  * When a primary channel has multiple sub-channels, we try to
1191  * distribute the load equally amongst all available channels.
1192  */
1193 struct vmbus_channel *vmbus_get_outgoing_channel(struct vmbus_channel *primary)
1194 {
1195         struct list_head *cur, *tmp;
1196         int cur_cpu;
1197         struct vmbus_channel *cur_channel;
1198         struct vmbus_channel *outgoing_channel = primary;
1199         int next_channel;
1200         int i = 1;
1201
1202         if (list_empty(&primary->sc_list))
1203                 return outgoing_channel;
1204
1205         next_channel = primary->next_oc++;
1206
1207         if (next_channel > (primary->num_sc)) {
1208                 primary->next_oc = 0;
1209                 return outgoing_channel;
1210         }
1211
1212         cur_cpu = hv_context.vp_index[get_cpu()];
1213         put_cpu();
1214         list_for_each_safe(cur, tmp, &primary->sc_list) {
1215                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1216                 if (cur_channel->state != CHANNEL_OPENED_STATE)
1217                         continue;
1218
1219                 if (cur_channel->target_vp == cur_cpu)
1220                         return cur_channel;
1221
1222                 if (i == next_channel)
1223                         return cur_channel;
1224
1225                 i++;
1226         }
1227
1228         return outgoing_channel;
1229 }
1230 EXPORT_SYMBOL_GPL(vmbus_get_outgoing_channel);
1231
1232 static void invoke_sc_cb(struct vmbus_channel *primary_channel)
1233 {
1234         struct list_head *cur, *tmp;
1235         struct vmbus_channel *cur_channel;
1236
1237         if (primary_channel->sc_creation_callback == NULL)
1238                 return;
1239
1240         list_for_each_safe(cur, tmp, &primary_channel->sc_list) {
1241                 cur_channel = list_entry(cur, struct vmbus_channel, sc_list);
1242
1243                 primary_channel->sc_creation_callback(cur_channel);
1244         }
1245 }
1246
1247 void vmbus_set_sc_create_callback(struct vmbus_channel *primary_channel,
1248                                 void (*sc_cr_cb)(struct vmbus_channel *new_sc))
1249 {
1250         primary_channel->sc_creation_callback = sc_cr_cb;
1251 }
1252 EXPORT_SYMBOL_GPL(vmbus_set_sc_create_callback);
1253
1254 bool vmbus_are_subchannels_present(struct vmbus_channel *primary)
1255 {
1256         bool ret;
1257
1258         ret = !list_empty(&primary->sc_list);
1259
1260         if (ret) {
1261                 /*
1262                  * Invoke the callback on sub-channel creation.
1263                  * This will present a uniform interface to the
1264                  * clients.
1265                  */
1266                 invoke_sc_cb(primary);
1267         }
1268
1269         return ret;
1270 }
1271 EXPORT_SYMBOL_GPL(vmbus_are_subchannels_present);
1272
1273 void vmbus_set_chn_rescind_callback(struct vmbus_channel *channel,
1274                 void (*chn_rescind_cb)(struct vmbus_channel *))
1275 {
1276         channel->chn_rescind_callback = chn_rescind_cb;
1277 }
1278 EXPORT_SYMBOL_GPL(vmbus_set_chn_rescind_callback);